Transformer-based anomaly detection for satellite telemetry data

Time series anomaly detection can help identify serious issues in complex systems, and can potentially reduce the risk of failures or operational disruptions by providing advance warning. Over the past decades, several methods, ranging from out-of-limit techniques to machine learning models have been developed to automate anomaly detection for satellite telemetry data. In recent years, transformer-based architectures have demonstrated considerable success in the problem of time series anomaly detection. In this paper, we present and compare the performance of various transformer architectures in detecting anomalies in satellite telemetry data, including the recently published ESA OPS-SAT telemetry dataset, and show how these architectures outperform the benchmarks conducted on this dataset. • Present the most advanced Transformer-based models for time series anomaly detection. • Compare the efficacy of various Transformer-based architectures in detecting anomalies in ESA OPS-SAT satellite telemetry data. • Provide a comprehensive and realistic assessment of the Transformer-based models in processing and identify anomalies.